Vous êtes sur la page 1sur 10

Matsuoka, M. and Estrada, M.


Development of Earthquake-Induced Building Damage

Estimation Model Based on ALOS/PALSAR Observing
the 2007 Peru Earthquake
Masashi Matsuoka∗ and Miguel Estrada∗∗
∗ InterdisciplinaryGraduate School of Science and Engineering, Tokyo Institute of Technology
Nagatsuta 4259-G3-2, Midori-ku, Yokohama 226-8502, Japan
E-mail: matsuoka.m.ab@m.titech.ac.jp
∗∗ Japan-Peru Center for Earthquake Engineering and Disaster Mitigation (CISMID), National University of Engineering

Av. Túpac Amaru 1150, Lima 25, Peru

E-mail: estrada@uni.edu.pe
[Received November 11, 2012; accepted December 14, 2012]

With the aim of developing a model for estimating which allows us to visually interpret the shapes of build-
building damage from synthetic aperture radar (SAR) ings and the states of large-scale damage [4, 5]. If a dis-
data in the L-band, which is appropriate for Peru, we aster strikes an extensive area, however, it takes consider-
propose a regression discriminant function based on able time to assess damage through visual interpretation,
field survey data in Pisco, which was seriously dam- making it impractical for rapid assessment. In addition,
aged in the 2007 Peru earthquake. The proposed because damage estimation by visual interpretation inher-
function discriminates among damage ranks corre- ently involves variability attributable to differences in the
sponding to the severe damage ratio of buildings using perception of the interpreter, there are problems with stan-
ALOS/PALSAR imagery of the disaster area before dardization and the general versatility of the method. In
and after the earthquake. By calculating differences in order to overcome such disadvantages, attempts have in-
and correlations of backscattering coefficients, which volved the use of a network of many volunteers to visually
were explanatory variables of the regression discrim- interpret building damage in an extensive disaster area,
inant function, we determined an optimum window namely, crowdsourcing [6]. There is a need to improve
size capable of estimating the degree of damage more the accuracy of the visual interpretation of damage. To
accurately. A normalized likelihood function for the this end, engineers with knowledge and experience in in-
severe damage ratio was developed based on discrim- terpreting aerial imagery should prepare instruction man-
inant scores of the regression discriminant function. uals that allow nonexperts to participate in this visual in-
The distribution of the severe damage ratio was accu- terpretation of damage, and should also establish a screen-
rately estimated, furthermore, from PALSAR imagery ing technique to eliminate unreliable data. Such a system
using data integration of the likelihood function with of exploiting the enormous human resources outside the
fragility functions in terms of the seismic intensity of disaster area is expected to become a technology used in
the earthquake. responding to major disasters.
Synthetic aperture radar (SAR), a type of “remote sens-
Keywords: severe damage ratio, ALOS/PALSAR, the ing sensor,” observes the surface of the earth day and
2007 Peru earthquake, likelihood function, backscattering night, regardless of weather. If the visual interpreta-
coefficient, data integration tion of damage from SAR imagery is practicable, it will
complement visual interpretation from optical sensor im-
agery. Unlike optical sensor images, which look like
1. Introduction photographs, however, SAR imagery represents the in-
tensity of microwave backscattering from the ground sur-
Satellite remote sensing is being increasingly used for face and is unfamiliar to nonexperts. It is therefore dif-
quick assessment of the impact of natural disasters oc- ficult to visually interpret damage from SAR images by
curring worldwide [1]. The size and location of affected crowdsourcing. It is thus expected that damage from
areas are estimated, for example, by comparing post- with SAR imagery will be extracted by computer-based image
pre-event imagery, since satellites orbiting the earth often processing [7–10]. Estimation models for building dam-
observe pre-event states at many locations [2, 3]. In recent age ratios have therefore been proposed [11–13] based
years, the resolution of optical sensors on board satellites on C-band (wavelength: 5.7 cm) SAR imagery in which
has increased to approximately 60-cm ground resolution, building damage data obtained from detailed field surveys
conducted after the Kobe earthquake in 1995 were used
1. This paper is translated with revision from the paper published in the as ground-truth data and the applicability of one of the
Journal of Japan Association for Earthquake Engineering, Vol.12, No.6, models used for damage extraction in other earthquakes
pp. 36-49 in Japanese.

346 Journal of Disaster Research Vol.8 No.2, 2013

Development of Earthquake-Induced Building Damage Estimation Model
Based on ALOS/PALSAR Observing the 2007 Peru Earthquake

Fig. 1. PALSAR imagery obtained before and after the 2007 Peru earthquake: (a) July 12, 2007; (b) August 27, 2007.

occurring in various countries and regions was investi- 2. PALSAR Images and Field Survey Data
gated [12]. The model was further improved so that it was
also applicable to imagery obtained by JERS-1/SAR and 2.1. Indices Obtained from PALSAR Imagery and
ALOS/PALSAR (PALSAR imagery), which is L-band Image Processing
(wavelength: 23 cm) SAR mounted on Japanese satel- On August 15, 2007, an earthquake measuring M8.0
lites. It was then applied to PALSAR imagery obtained in with an epicenter 40 km northwest of Chincha, Peru.
the 2007 Peru earthquake and the 2008 Wenchuan, China, The city of Pisco in the Ica Region and the surround-
earthquake [13]. It was found from comparison with dam- ing area were devastated by the earthquake, which left
age assessment reports of these earthquakes, etc., that more than 500 people dead or missing and completely de-
local damage areas could not be detected because the stroyed more than 35,000 buildings. About 2 weeks after
model was based on the ground-truth data of the 1995 the earthquake, high-resolution PALSAR imagery of the
Kobe earthquake. In other words, it was demonstrated coastal area was obtained. Figs. 1(a) and (b) shows im-
that application of the model to other countries and re- ages obtained on July 12, 2007, before the earthquake and
gions, which have urban structures, building types, and on August 27, 2007, after the quake. The nominal ground
damage situations different from Japan, gave inaccurate resolution was approximately 10 m and the image pixel
results [13]. size 12.5 m.
Following procedures reported in our previous pa- Two indices – the difference between post- and pre-
pers [11, 13], this paper develops a model for estimating earthquake images and the correlation coefficients of
a severe damage ratio of buildings that reflects the build- backscattering coefficients – were calculated from pre-
ing types and damage situation in Peru. The model is an and post-earthquake PALSAR images. Following the ac-
estimation model for severe damage ratio optimized for curate positioning of both pre- and post-event images, a
Peru because it is based on PALSAR imagery of Pisco speckle reduction filter was applied to each image [14],
after the 2007 Peru earthquake and on building damage then differences and correlation coefficients were calcu-
data obtained from field surveys. Specifically, an opti- lated from Eqs. (1) and (2), below. The difference is ob-
mum window size for image processing was determined tained by subtracting the average value of the backscat-
and a normalized likelihood function for the severe dam- tering coefficient within an N × N pixel window of the
age ratio was derived. The model was used, furthermore, pre-event image from that of the post-event image. The
to estimate damage by integrating image data with seis- correlation coefficient is also calculated from the same
mic intensity data and comparing results with the actual N × N pixel window. Analysis results from the 1995
damage to be checked. Great Hanshin-Awaji Earthquake Disaster (widely called
the Kobe earthquake) showed that differences (after – be-
fore) yielded negative values, with the spatial distribu-
tion of backscattering coefficients decreasing with build-
ing damage and as compared with that in pre-event im-

Journal of Disaster Research Vol.8 No.2, 2013 347

Matsuoka, M. and Estrada, M.

Fig. 2. Building damage data at Pisco based on a field survey. (a) Damage level by lot (Estrada et al., 2008). (b) Distribution of the
severe damage ratio.

agery, resulting in an overall decrease in correlation coef- Table 1. Range of severe damage ratio and median values
for damage ranks.
ficient [10].
d = 10 · log10 Ia
¯ i − 10 · log10 Ib
¯ i . . . . . . (1) Damage Rank
Severe Damage
Median Value (%)
Ratio D (%)
r= C1 D=0 0.0
N N N C2 0 < D < 6.25 3.13
N ∑ Iai Ibi − ∑ Iai ∑ Ibi
i=1 i=1 i=1 C3 6.25 ≤ D < 12.5 9.38

  2    2  C4 12.5 ≤ D < 25 18.75
 N N N N
 N ∑ Ia2 − ∑ Iai · N ∑ Ib2i − ∑ Ibi C5 25 ≤ D < 50 37.5
i=1 i=1 i=1 i=1
C6 50 ≤ D ≤ 100 75.0
. . . . . . . . . . . . . . . . . (2)
where d represents the difference in backscattering co-
efficients [dB], r is the correlation coefficient, and N is damaged. Lots and buildings were mostly in a one-to-
the number of pixels within the window to be calculated. one correspondence. For multiple buildings located on
Iai and Ibi represent the i th pixel values of post- and pre- large lots, information on the most damaged building was
event images, respectively, and Ia ¯ i and Ib
¯ i represent aver- recorded. Fig. 2(a) shows the distribution of damage lev-
age values of N × N pixels surrounding the i th pixel. els by lot based on the field survey in Pisco.
Damage levels were classified into the following four:
Grave (Serious), Severo (Severe), Leve (Slight), and Sin
2.2. Severe Damage Ratio of Buildings Based on daño (No damage). Lots that could not be investigated
Field Survey Data were categorized as “Not available.” Grave corresponds
The target area of the damage estimation model is the to G5 in the classification of the European Macroseismic
city of Pisco. Damage data used in this study were col- Scale (EMS-98) [16], Severo to G4 and G3, Leve to G2,
lected by members of the Japan-Peru Center for Earth- and Sin daño to G1. The damage ratio was calculated
quake Engineering and Disaster Mitigation (CISMID), based on these GIS data. In order to take sizes of lots, in-
National University of Engineering, Peru. CISMID per- cluding vacant lots, into consideration and to calculate re-
sonnel performed detailed on-site investigations of build- liable damage ratios, the city of Pisco was split into a grid
ings in more than 10,000 lots just after the earthquake [15] of 3.75 × 3.75 arc-seconds (approximately 120-m grid)
and this was considered to be the most reliable data on and estimation was performed only on grids with 10 or
the Peru earthquake. Investigation items included build- more lots. The severe damage ratio of buildings in a grid
ing lot codes, building use, structure types, floor number, was calculated as the ratio of the number of Grave to the
and damage level, all of which were combined with ge- total number of buildings in the grid. The severe damage
ographic information system (GIS) data. Approximately ratio was classified into the following six damage ranks:
97% of buildings in the area were masonry structures – C1, 0% severe damage ratio in a grid; C2, more than 0%
18% adobe structures and 79% burnt brick structures – and less than 6.25%; C3, 6.25% or more and less than
which are comparatively weak and are prone to collapse in 12.5%; C4, 12.5% or more and less than 25%; C5, 25%
general. In the earthquake, adobe buildings were greatly or more and less than 50%; and C6, 50% or more. Ta-

348 Journal of Disaster Research Vol.8 No.2, 2013

Development of Earthquake-Induced Building Damage Estimation Model
Based on ALOS/PALSAR Observing the 2007 Peru Earthquake

ble 1 shows the correspondence of damage rank, severe

damage ratio and median values. The distribution of the
severe damage ratio is shown in Fig. 2(b).

3. Estimation of Severe Damage Ratio Based

on Regression Discriminant Function

3.1. Influence of Window Size on Accuracy of Dam-

age Discrimination

In previous studies [10–13], a speckle reduction filter

with a 21 × 21 pixel window was applied to SAR images
and differences and correlation coefficients were calcu-
lated from a 13 × 13 pixel window. Although these win-
dow sizes suited extraction of the damaged area from 30-
m resolution SAR imagery of an area affected by the Kobe
earthquake and field survey data in the Hanshin area [10], Fig. 3. Relationship between the size of speckle reduction
it was uncertain whether these window sizes would be the filters and the correlation ratio.
optimum for extracting damage in a city in Peru, where
building types are different from those in the Hanshin
area, with approximately 10-m resolution, which is the
same as PALSAR imagery resolution. Accordingly, the
change in the accuracy of damage discrimination was ex-
amined by varying speckle reduction filter size and calcu-
lation window size for Pisco data sets.
The influence of speckle reduction filters was examined
first. A Lee filter size [14] that is variable from 3 × 3 to
51 × 51 pixels was applied to pre- and post-event images
and differences and correlation coefficients were calcu-
lated based on Eqs. (1) and (2). Images of differences
and correlation coefficients were overlaid on field survey
data and 800 pixels were randomly extracted from areas
corresponding to each of the six damage ranks shown in
Table 1 (4,800 pixels in total) to create a training sam-
ple. For quantitative evaluation of the severe damage ra-
tio, regression discriminant analysis [17] – a method of
multiple-group discriminant analysis that uses differences
and correlation coefficients of the six damage ranks – was Fig. 4. Relationship between the calculation window size
applied. Window sizes of 7 × 7, 13 × 13, and 21 × 21 pix- and the correlation ratio.
els were examined to calculate differences and correlation
coefficients. Fig. 3 shows the correlation ratio of regres-
sion discriminant functions representing the ability to dis- shows correlation ratio against pixel dimension calcu-
criminate six damage ranks against the pixel dimension lated from window size. Window size increased from
calculated from the size of the speckle reduction filter. In 3 × 3, so the correlation ratio increased, reaching a limit
this figure, a larger correlation ratio means better discrim- at around 13 × 13 pixels. This was also found for data
inant ability of damage ranks. The relationship between sets in the Hanshin area before and after the Kobe earth-
pixel dimension and correlation ratio is slightly compli- quake [10]. The reasons why the correlation ratio in-
cated because as the size of the filter increases, the corre- creases includes the fact that damaged building groups
lation ratio decreases, but turns upward at 15 × 15 pixels. spread to some extent and, in addition, backscattering of
The correlation ratio obtained when the filter size is in- each damaged building has a spatial extent. Interestingly,
creased to the largest one of 41 × 41 pixels is almost the although the pixel size of SAR images in the Hanshin
same as that obtained without any filter, however, so it area and Pisco are different, the window size at which
was determined that no filter would be used in this study. the correlation ratio reached a limit was the same, i.e.,
The influence of window size on the accuracy of dis- at around 13 × 13 pixels. Although this is considered to
crimination was examined next. The correlation ratio arise from complex factors such as the fact that different
of regression discriminant functions was calculated us- damage situations are involved, the detail remains a chal-
ing varying window size from 3 × 3 to 51 × 51 pixels to lenge to be addressed. It should be noted that although
calculate differences and correlation coefficients. Fig. 4 the correlation ratio reached a maximum in Fig. 4 when

Journal of Disaster Research Vol.8 No.2, 2013 349

Matsuoka, M. and Estrada, M.

Fig. 6. Normal frequency distribution model of discriminant

score ZRp .

Table 2. Average values and standard deviations in the like-

Fig. 5. Relationship between differences in backscattering lihood function of data in PALSAR intensity imagery.
coefficients and correlation coefficients for damage ranks.
Damage Rank Average of ZRp Standard Deviation
C1 -1.470 0.323
21 × 21 pixels were used, the value is almost the same as C2 -1.355 0.291
that for 13 × 13 pixels. A larger window size, however,
C3 -1.332 0.281
does not offer the capability for detecting small spatial
changes within a window, so we decided in our study to C4 -1.200 0.331
use a 13 × 13 pixel window, which was also used in the C5 -1.052 0.415
previous study. C6 -0.887 0.484

3.2. Derivation of Regression Discriminant Func-

tion and Likelihood Function means the probability of being in each damage rank when
To calculate difference d and correlation coefficient r, ZRp is given. Specifically, the frequency distribution of
a window size of 13 × 13 pixels was adopted as the opti- ZRp of 800 randomly extracted pixels from each damage
mum size to interpret damage in Pisco based on the anal- rank is modeled as a normal distribution. Fig. 6 shows
ysis of data sets, and no speckle reduction filter was used. normal distribution models (likelihood function). Table 2
Fig. 5 shows a scatter diagram by damage rank. The re- shows average values and standard deviations of ZRp for
gression discriminant function calculated from the two in- individual damage ranks. The higher the damage rank,
dices is shown in Eq. (3): the larger the discriminant score ZRp . Because distribu-
tion curves of some damage ranks cross in regions with
ZRp = −0.089 d − 2.576 r . . . . . . . . (3)
low discriminant scores, however, discrimination in areas
where ZRp represents the discriminant score derived from with low damage ranks is not possible. Fig. 7 shows nor-
PALSAR imagery. While coefficients of d and r in dis- malized likelihood functions in which the sum of the like-
criminant score ZR j derived from JERS-1/SAR imagery lihood of all damage ranks in Fig. 6 becomes 1.0. For re-
of the Kobe earthquake were −1.277 and −2.729, re- gions where ZRp is −2.2 or less, a constant value obtained
spectively [13], coefficients derived here from PALSAR by extrapolating the value at ZRp = −2.2 is used in order
imagery of Pisco were −0.089 and −2.576. These coeffi- to avoid reversing the sequence of the severe damage ra-
cients indicate the degree of influence of d or r on the dis- tio caused by distribution curves crossing. Average values
criminant score. Comparison between the Hanshin area and standard deviation of the estimated severe damage ra-
and Pisco demonstrated that the coefficient of r, the corre- tio against discriminant score ZRp are thus obtained from
lation coefficient, was almost the same in the two regions, median values of the damage rank in Table 1 and the dis-
but the coefficient of d, the difference in backscattering tribution shown in Table 2 and Fig. 7. Fig. 8 shows curves
coefficients, of Pisco was small, or approximately zero, of average values and average values ± standard deviation
compared with that of the Hanshin area. The influence of of the severe damage ratio estimated from ZRp . The severe
d on discrimination of the damage rank in Pisco is there- damage ratio increases with increasing ZRp . Because the
fore negligible. discriminant score was adjusted to make the constant term
Next, a likelihood function for estimating the severe zero, relative positions on the horizontal axis in Fig. 8 are
damage ratio from discriminant score ZRp is formed arbitrary. Taking this into account, the normalized like-
following procedures similar to those in the previous lihood function derived from field survey data and PAL-
study [11, 13]. The likelihood function in this study SAR imagery of Pisco gives a high severe damage ratio

350 Journal of Disaster Research Vol.8 No.2, 2013

Development of Earthquake-Induced Building Damage Estimation Model
Based on ALOS/PALSAR Observing the 2007 Peru Earthquake

Fig. 7. Normalized likelihood function of discriminant

score ZRp .

Fig. 9. Discriminant score ZRp estimated from PALSAR


Fig. 8. Relationship between the severe damage ratio (av-

erage values and standard deviations) and the discriminant
scores of Pisco ZRp and Hanshin ZR j .

(average value) in regions with low discriminant scores

compared with that derived from field survey data and
JERS-1/SAR imagery of the Hanshin area [13] (shown
together in Fig. 8). It also indicates that small changes
in backscattering characteristics have a large influence on Fig. 10. Severe damage ratio of buildings estimated from
the estimation of the severe damage ratio. This could be the normalized likelihood function (average values).
influenced by differences in building damage situations
between Pisco and the Hanshin area.
Figure 9 shows the ZRp distribution obtained from pre-
and post-earthquake PALSAR images, and Fig. 10 shows
the severe damage ratio (average values) estimated from
ZRp . ZRp values and the severe damage ratio are slightly
larger at the center of Pisco. It should be noted that the
target area is restricted to urban areas where the cardinal
effect can be expected, therefore, areas whose backscat-
tering coefficients are small (−5 dB or under) in pre-event
images are masked. The distribution of the severe damage
ratio of Pisco city estimated based on PALSAR imagery
agrees well with field survey data. Areas with high severe
damage ratios are found in sections of farmland, however,
because they are not adequately masked owing to varia-
tions in backscattering characteristics caused by vegeta- Fig. 11. Relationship between seismic intensity and average
tion. values and standard deviations in the fragility function.

Journal of Disaster Research Vol.8 No.2, 2013 351

Matsuoka, M. and Estrada, M.

Fig. 12. Estimated severe damage ratio obtained by data integration of PALSAR imagery and seismic intensity
data: (a) average values, (b) standard deviations.

4. Severe Damage Ratio Estimation by Integra- normalized likelihood of ZRp of each damage rank (equiv-
tion with Seismic Intensity Data alent to the probability of being in each damage rank when
ZRp is given), and then normalizing (that is, constraining
In a case where only PALSAR imagery is used and the sum of probabilities to be one) [11]. Figs. 12(a) and
ZRp < −2.2, distributions overlap as shown in Fig. 6 (b) shows average values and standard deviations in the
and contain only slightly more information than complete severe damage ratio based on the probability updated for
noninformation. In other words, the average severe dam- the 7 × 7 = 49 combinations of the two indices. Com-
age ratio was 13.0% with a standard deviation of 21.0, pared with Kobe earthquake results [13], seismic inten-
compared to an average severe damage ratio of 35.4% and sity data contribute to estimating the severe damage ratio
a standard deviation of 31.2% when an equal probability where ZRp of SAR imagery is low, so the severe damage
of 1/6 is given for each damage rank. As is shown in ratio increases steeply and the seismic intensity data con-
Fig. 10, a severe damage ratio of approximately 20% is tribution is nearly zero where SAR imagery ZRp is high.
estimated even in broadly undamaged areas. Seismic in- U.S. Geological Survey (USGS) ShakeMap [20] was
tensity data is therefore used as supplementary data for used for earthquake seismic intensity data. Peak Ground
highly accurate estimation in all regions, including re- Velocity distribution data from ShakeMap [21] was con-
gions of low severe damage ratio. A fragility function in verted by Fujimoto and Midorikawa [22] into measured
terms of seismic intensity data has been established based seismic intensity, and is shown superimposed on PAL-
on data obtained from the Kobe earthquake [11]. An im- SAR imagery in Fig. 13. Large shaking of 6− to 6+
proved fragility function that considers the fragility of on the Japan Meteorological Agency (JMA) seismic in-
masonry buildings in Peru has also been developed [13]. tensity scale is found all over the image. The result of
Fig. 11 shows the fragility function for Peru used in this estimating the severe damage ratio by data integrating
study and that for the Hanshin area for comparison. It discriminant score ZRp data obtained from PALSAR im-
corresponds roughly with previous fragility functions [18, agery (Fig. 9) and measured seismic intensities (Fig. 13)
19]; the severe damage ratio of buildings in Peru becomes is shown in Fig. 14(a). An enlarged view of the Pisco area
higher than that in the Hanshin area if buildings are af- is shown in Fig. 14(b). Because built-up area is the target,
fected by earthquakes of the same seismic intensity. areas with backscattering coefficients of −5 dB or less
In line with the work of Nojima et al. [11], Bayesian are masked. Fig. 15 shows enlarged views in the Pisco
updating theory has been used to improve the accuracy area of (a) distribution of the severe damage ratio inves-
of estimating the severe damage ratio. It does so by inte- tigated by the field survey, (b) the same distribution es-
grating PALSAR imagery ZRp with seismic intensity data. timated from PALSAR imagery alone, and (c) the same
Specifically, the probability of being in each damage rank distribution based on the data integration of PALSAR im-
is updated by multiplying the probability of being in each agery and seismic intensity data. Distribution of the se-
damage rank when seismic intensity data is given with the vere damage ratio estimated based on the data integration

352 Journal of Disaster Research Vol.8 No.2, 2013

Development of Earthquake-Induced Building Damage Estimation Model
Based on ALOS/PALSAR Observing the 2007 Peru Earthquake

Fig. 13. Distribution of measured seismic intensity esti-

mated from ShakeMap.

Fig. 15. Comparison of distribution maps of the severe dam-

age ratio in Pisco (enlarged views): (a) field survey, (b) esti-
mation based on PALSAR imagery alone, and (c) estimation
based on the data integration of PALSAR imagery and seis-
Fig. 14. Estimated severe damage ratio obtained by data mic intensity.
integration of PALSAR imagery and seismic intensity data
(average values).

lished in this study, however, could extract damage in this

area. This fact suggests that, in order to accurately esti-
of PALSAR imagery and the seismic intensity data is in mate damage areas, it is important to use a damage ex-
good agreement with the distribution based on the field traction model suited to the target area because there are
survey, especially in areas with a high severe damage ra- difference in urban structures, building types, and damage
tio in the central part of Pisco. looking at a broad area situations between Japan and Peru.
all over Pisco, the area overestimated by estimation based
on PALSAR imagery alone could be reestimated appro-
priately. That is, it would be estimated at the level of 5. Conclusions
the actual severe damage ratio by using estimation based
on data integration with seismic intensity data. When the To develop a technology to quickly assess areas af-
model based on the Kobe earthquake was applied to Pisco fected by earthquakes using imagery of L-band synthetic
without improvement, damage in the western coastal area aperture radar (SAR) mounted on satellites, an estima-
of Pisco could not be detected [13]. The model estab- tion model of the severe damage ratio of buildings, op-

Journal of Disaster Research Vol.8 No.2, 2013 353

Matsuoka, M. and Estrada, M.

timized for Peru, has been created based on field survey China, earthquake based on visual detection of satellite optical im-
data and PALSAR imagery of the city of Pisco struck by ages,” Journal of Japan Association for Earthquake Engineering,
Vol.10, No.3, pp. 46-57, 2010 (in Japanese with English abstract).
the 2007 Peru earthquake. Regression discriminant anal- [5] S. Matsuzaki, F. Yamazaki, M. Estrada, and C. Zavala, “Visual
ysis has been performed in which explanatory variables, damage interpretation of buildings using QuickBird images follow-
ing the 2007 Pisco, Peru earthquake,” Journal of Institute of Social
namely differences and correlation coefficients, were cal- Safety Science, No.13, pp. 407-413, 2010 (in Japanese with English
culated from pre- and postearthquake PALSAR images. abstract).
[6] S. Ghosh, C. K. Huyck, M. Greene, S. P. Gill, J. Bevington, W.
Target groups were six damage ranks classified based on Svekla, R. DesRoches, and R. T. Eguchi, “Crowdsourcing for rapid
the severe damage ratio of buildings. Examination of an damage assessment: the Global Earth Observation Catastrophe
optimum window size for calculation of the differences Assessment Network (GEO-CAN),” Earthquake Spectra, Vol.27,
No.S1, pp. S179-S198, 2011.
and correlation coefficients have resulted in the same pa- [7] G. A. Arciniegas, W. Bijker, N. Kerle, and V. A. Tolpekin,
rameter as that obtained from data from the 1995 Kobe “Coherence- and amplitude-based analysis of seismogenic damage
in Bam, Iran, using Envisat ASAR data,” Transactions on Geo-
earthquake. A regression discriminant function has been science and Remote Sensing, Vol.45, pp. 1571-1581, 2007.
derived that considers urban structure, building type, the [8] S. Midorikawa and H. Miura, “Extraction of landslide areas due
damage situation in Peru, and the pixel resolution of PAL- to the 2008 Iwate-Miyagi-nairiku, Japan earthquake from high-
resolution SAR image,” Journal of Japan Association for Earth-
SAR imagery. Furthermore, normalized likelihood func- quake Engineering, Vol.10, No.3, pp. 25-32, 2010 (in Japanese with
tions for the severe damage ratio have been obtained from English abstract).
[9] Y. Ito and M. Hosokawa, “An estimation model of damage degree
the discriminant score of the regression discriminant func- using interferometric SAR data,” IEEJ Transactions on Electronics,
tion. It has also been demonstrated that the distribution of Information and Systems, Vol.122-C, No.4, pp. 617-623, 2002 (in
Japanese with English abstract).
the severe damage ratio could be estimated at high accu- [10] M. Matsuoka and F. Yamazaki, “Use of satellite SAR intensity im-
racy based on data integration of PALSAR imagery with a agery for detecting building areas damaged due to earthquakes,”
fragility function based on the seismic intensity data of the Earthquake Spectra, Vol.20, No.3, pp. 975-994, 2004.
[11] N. Nojima, M. Matsuoka, M. Sugito, and K. Ezaki, “Quantitative
Peru earthquake. Similar to PALSAR, an L-band SAR, estimation of building damage based on data integration of seismic
called PALSAR-2 will be mounted on a satellite, ALOS- intensities and satellite SAR imagery,” Journal of JSCE, Division A,
Vol.62, No.4, pp. 808-821, 2006 (in Japanese with English abstract).
2, to be launched in 2013. If an earthquake occurs in Peru [12] M. Matsuoka and F. Yamazaki, “Comparative analysis for detecting
in the future, the model proposed in this study could be areas with building damage from several destructive earthquakes
used for damage assessment. By preparing a highly reli- using satellite synthetic aperture radar images,” Journal of Applied
Remote Sensing, SPIE, Vol.4, 041867, 2010.
able fragility function that fits the actual situation of build- [13] M. Matsuoka and N. Nojima, “Building damage estimation by inte-
ings in Peru and integrating this with seismic intensity gration of seismic intensity information and satellite L-band SAR
imagery,” Remote Sensing, MDPI, Vol.2, No.9, pp. 2111-2126,
data based on strong-motion observation and simulation, 2010.
we will be able to assess damage more accurately. [14] J. S. Lee, “Digital image enhancement and noise filtering by use of
local statistics,” IEEE Trans. Pattern Analysis and Machine Intelli-
gence, Vol.2, No.2, pp. 165-168, 1980.
[15] M. Estrada, C. Zavala, and Z. Aguilar, “Damage study of the Pisco,
Acknowledgements Peru earthquake using GIS and satellite images,” Proceedings of
International Workshop for Safer Housing in Indonesia and Peru,
PALSAR data are the property of Japan’s Ministry of Economy, 2008.
Trade and Industry (METI) and the Japan Aerospace Exploration [16] G. Grünthal, “European Macroseismic Scale 1998 (EMS-98),” Eu-
Agency (JAXA) and have been processed using the GEO Grid of ropean Seismological Commission, 1998.
the National Institute of Advanced Industrial Science and Tech- [17] T. Okuno, H. Kume, T. Haga, and T. Yoshizawa, “Multivariate
statistical methods,” Union of Japanese Scientists and Engineers,
nology (AIST). This study was supported in part by a Science Tokyo, Japan, pp. 259-321, 1981 (in Japanese).
and Technology Research Partnership for Sustainable Develop- [18] A. Coburn and R. J. S. Spence, “Earthquake protection – chapter 9
ment (SATREPS) project titled Enhancement of Earthquake and earthquake risk modeling –,” John Wiley & Sons, p. 436, 2002.
Tsunami Disaster Mitigation Technology in Peru (Principal In- [19] K. A. Porter, K. S. Jaiswal, D. J. Wald, M. Greene, and C. Comartin,
“WHE-PAGER project: a new initiative in estimating global build-
vestigator: Prof. Fumio Yamazaki); a New Energy and Industrial
ing inventory and its seismic vulnerability,” Proc. The 14th World
Technology Development Organization (NEDO) Industrial Tech- Conference on Earthquake Engineering, No.S23-016, 2008.
nology Research Grant Program, titled Development of realtime [20] US Geological Survey, “ShakeMap,” http://earthquake.usgs.gov/
tsunami damage detection technology for expeditious disaster re- eqcenter/shakemap/ [accessed February 2012]
sponse of Japan and ASEAN countries (Leader: Prof. Shunichi [21] US Geological Survey, “ShakeMap us2007gbcv,” http://earthquake.
usgs.gov/eqcenter/shakemap/global/shake/2007gbcv/ [accessed
Koshimura); and a grant in aid for scientific research (Research June 2009]
No. 21310119; Principal Investigator: Prof. Fumio Yamazaki). [22] K. Fujimoto and S. Midorikawa, “Empirical method for estimating
We would like to express our gratitude for these contributions. J.M.A. instrumental seismic intensity from ground motion parame-
ters using strong motion records during recent major earthquakes,”
Journal of Institute of Social Safety Science, No.7, pp. 214-246,
2005 (in Japanese with English abstract)
[1] International Charter Space and Major Disaster, http://www. disas-
terscharter.org [accessed February 2012]
[2] P. Gamba, A. Marazzi, and E. Costamagna, “Satellite data analysis
for earthquake damage assessment,” Proc. of SPIE 3222, Earth Sur-
face Remote Sensing (G. Cecchi, E. T. Engman, E. Zilioli (Eds.)),
pp. 340-350, 1997.
[3] K. Saito, R. J. S. Spence, C. Going, and M. Markus, “Using high-
resolution satellite images for post-earthquake building damage as-
sessment: A study following the 26 January 2001 Gujarat earth-
quake,” Earthquake Spectra, Vol.20, No.1, pp. 145-169, 2004.
[4] H. Miura and S. Midorikawa, “Distribution of building damage in
the southeastern part of Beichuan county due to the 2008 Sichuan,

354 Journal of Disaster Research Vol.8 No.2, 2013

Development of Earthquake-Induced Building Damage Estimation Model
Based on ALOS/PALSAR Observing the 2007 Peru Earthquake

Name: Name:
Masashi Matsuoka Miguel Estrada

Affiliation: Affiliation:
Associate Professor, Department of Built Envi- General Director, CISMID
ronment, Tokyo Institute of Technology Associate Professor, Faculty of Civil Engineer-
ing, National University of Engineering

Address: Address:
Nagatsuta 4259-G3-2, Midori-ku, Yokohama 226-8502, Japan Av. Tupac Amaru 1150, Rimac, Lima, Peru
Brief Career: Brief Career:
1992 Research Associate, Tokyo Institute of Technology 1998-2000 Master of Engineering in the field of Civil Engineering, The
1996 Engineer, Remote Sensing Technology Center of Japan University of Tokyo
1998 Deputy Team Leader, RIKEN 2000-2004 Ph.D. of Civil Engineering, The University of Tokyo
2004 Team Leader, National Research Institute for Earth Science and 2004-present Associate Professor, Faculty of Civil Engineering, National
Disaster Prevention University of Engineering
2007 Senior Research Scientist, National Institute of Advanced Industrial 2013-present General Director, Japan-Peru Center for Earthquake
Science Technology Engineering Research and Disaster Mitigation (CISMID), Faculty of Civil
2010 Division Chief, National Institute of Advanced Industrial Science Engineering, National University of Engineering
Technology Selected Publications:
2012- Associate Professor, Tokyo Institute of Technology • M. Estrada, H. Miura, F. Yamazaki, and S. Midorikawa, “Evaluation of
Selected Publications: Social Seismic Vulnerability through High Resolution Satellite Imagery,”
• Matsuoka and Yamazaki, “Use of Satellite SAR Intensity Imagery for 15th World Conference on Earthquake Engineering, Portugal, 2012.
Detecting Building Areas Damaged due to Earthquakes,” Earthquake • M. Estrada, C. Zavala, and Z. Aguilar, “Use of Geomatics for Disaster
Spectra, EERI, Vol.20, No.3, pp. 975-994, 2004. Management – Case Study 2007 Peru, Pisco Earthquake,” 7th International
• Matsuoka and Nojima, “Building Damage Estimation by Integration of Workshop on Remote Sensing and Disaster Response, USA, 2009.
Seismic Intensity Information and Satellite L-band SAR Imagery,” Remote • M. Estrada, M. Matsuoka, and F. Yamazaki, “Use of Optical Satellite
Sensing, MDPI, Vol.2, No.9, pp. 2111-2126, 2010. Images for the Recognition of Areas Damaged by Earthquakes” 6th
• Matsuoka and Yamazaki, “Comparative Analysis for Detecting Areas International Conference on Seismic Zonation, USA, 2000.
with Building Damage from Several Destructive Earthquakes Using Academic Societies & Scientific Organizations:
Satellite Synthetic Aperture Radar Images,” Journal of Applied Remote • Peruvian Board of Engineers
Sensing, SPIE, Vol.4, 041867, 2010. • Earthquake Engineering Research Institute
Academic Societies & Scientific Organizations:
• Earthquake Engineering Research Institute (EERI)
• Architectural Institute of Japan (AIJ)
• Remote Sensing Society of Japan (RSSJ)

Journal of Disaster Research Vol.8 No.2, 2013 355